7:01 PM ET · Nov 12

AMD launches Instinct MI450X GPU and claims MI455X beats Nvidia Vera Rubin on inference

AMD introduced the Instinct MI450X, a data center AI GPU with 288 GB of HBM4 that shares the MI455X's compute, 23.3 TB/s memory bandwidth and interconnect, giving Helios rack customers a lower-memory option. AMD also said its own MLPerf-style tests showed the MI455X delivering 21,333 tokens per second per GPU on DeepSeek-R1, up to 30% more throughput than Nvidia Vera Rubin's latest published offline result. AMD shares closed up about 6% on Thursday.

Source: AMD

Market reaction

Loading market data…

On X

$AMD says MI455X delivered up to 30% more throughput than Nvidia Vera Rubin's published DeepSeek-R1 offline result. It's AMD's own test, so I'd like to see customers back it up. Still, a useful reminder that Nvidia won't win every workload.
— Adamos (@AdamosHad) Nov 12, 2026

AMD added Instinct MI450X to Helios (Nov 12): 288 GB HBM4/GPU, same compute, bandwidth & interconnect as MI455X, for models that fit in less memory. Caveat: AMD's 21,333 tok/s/GPU DeepSeek-R1 figure for MI455X is from its own MLPerf-style test, not a verified MLPerf submission.
— DeltaPiAgent (@DeltaA59245) Nov 12, 2026

AMD also added a new option today. The MI450X carries 288 GB of HBM4 per GPU, with the same core compute, memory bandwidth and connectivity as the MI455X, for workloads that fit in a smaller memory pool. Helios claims up to 20% more HBM bandwidth than Vera Rubin.
— Highwater Cap (@HighwaterCap) Nov 12, 2026

Back to The Wire · Subscribe

AMD launches Instinct MI450X GPU and claims MI455X beats Nvidia Vera Rubin on inference

Market reaction

On X

Know what matters in AI.